Machine learning algorithms with body fluid parameters: an interpretable framework for malignant cell screening in cerebrospinal fluid.

Objectives: This study aimed to develop and validate a machine learning (ML) model utilizing cerebrospinal fluid (CSF) body fluid parameters from hematology analyzers to screen for malignant cells.

Methods: We analyzed 643 consecutive CSF samples from patients with central nervous system symptoms, with 191 samples classified as positive for malignant cells based on cytological examination, for model derivation. Body fluid parameters were measured using the body fluid mode of a hematology analyzer. Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to identify predictive biomarkers, followed by performance evaluations of six ML algorithms. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). The selected model was also externally validated with an additional 136 CSF samples.

Results: The median leukocyte (WBC) and total nucleated cell (TNC) counts in the cytology-positive samples were significantly lower than those in the cytology-negative samples (5.4 vs. 31.8 and 7.4 vs. 32.6, respectively, p<0.001). The support vector machine (SVM) model achieved the highest area under the curve (AUC) of 0.899 (SD: 0.035) and the highest sensitivity of 0.827 (SD: 0.059) in internal validation. SHAP analysis identified the percentage of high fluorescence cells and monocytes as the two most significant predictors, both positively correlated with malignant cell outcomes. External validation demonstrated a comparable AUC and sensitivity, confirming the model's generalizability.

Conclusions: We developed an ML model that predicts cytological outcomes in CSF using routinely available body fluid parameters. The model demonstrated consistent performance during external validation.